Intelligent Narrative Technologies: Papers from the 2012 AIIDE Workshop
AAAI Technical Report WS-12-14
Gestural Interactions for Interactive Narrative Co-Creation
Andreya Piplica, Chris DeLeon, and Brian Magerko
Georgia Institute of Technology
{apiplica3, cdeleon3, magerko}@gatech.edu
Abstract
convey dialogue. Humans, on the other hand, are typically
limited to interacting through mouse and keyboard, which
do not afford much physical expression. This disparity in
interaction capabilities has been informally discussed in
the community as the human puppet problem.
The two main paradigms for human interaction in
interactive narratives are language-based and menu-based
interaction. With language-based interaction, users type or
speak natural language utterances to interact with virtual
characters (Mateas and Stern 2003; Riedl and Stern 2006;
Aylett et al. 2005). With menu-based interaction, users
select from dialogue or actions in a visual list (McCoy et
al. 2011). Some interactive narratives respond to the user’s
actions in a game world, which are still mediated by
keyboard-and-mouse input modalities (Magerko and Laird
2003; Young 2001). None of these approaches involves
full-bodied gestural interaction directly from the user.
Embodied agents can add animated gestures to a user’s
input (Zhang et al. 2007; El-Nasr et al. 2011), but this does
not allow the user to communicate with their own gestures.
Fully immersive environments (Hartholt, Gratch, and
Weiss 2009) allow users to interact with virtual characters
and objects as they would real people and objects, but
often focus on established stories rather than co-creating
new ones.
Improvisational theatre (improv) provides a real-world
analogue to embodied co-creative interactive narrative
experiences (J. Tanenbaum and Tanenbaum 2008;
Magerko and Riedl 2008). Improvisational actors
collaboratively create novel stories in real-time as part of a
performance. They communicate with each other through
both dialogue and full-body movements. Movements can
portray characters (including their status, mood, and intent)
(Laban and Ullmann 1971) as well as contribute actions to
a scene which establish the activity or location (Johnstone
1999). All communications occur within the (diegetic)
context of an improvised scene; therefore, improvisers
cannot use explicit communication to resolve conflicting
ideas about the scene they are creating (Magerko et al.
2009). In an ideal real-time co-creative system, human
This paper describes a gestural approach to interacting with
interactive narrative characters that supports co-creativity.
It describes our approach using a Microsoft Kinect to
created a short scene with an intelligent avatar and an AIcontrolled actor. It describes our preliminary user studies
and a recommendation for future evaluation.
Introduction
One under-researched approach to interactive narrative
seeks to allow humans and intelligent agents to co-create
narratives in real-time as equal contributors. In an ideal
narrative co-creation experience, a human interactor and an
intelligent agent act as characters in a novel (i.e. not prescripted) narrative. The experience is valuable both as a
means of creating a narrative and engaging in a
performance. In such a system, the story elements for an
experience are not pre-defined by the AI’s knowledge (e.g.
a drama manager who is imbued with the set of possible
story events for an experience), but rather the human and
AI have a) similar background knowledge and b) similar
procedural knowledge for how to collaboratively create a
story together. While this stance is within the boundaries of
emergent narrative (i.e. approaches to interactive narrative
that do not have prescribed story events), the typical
approach in emergent narrative systems relies on strongly
autonomous systems (i.e. rational agents pursuing their
goals to elicit story events), which may not engage directly
with a user and their narrative goals. Co-creation describes
a more collaborative process between a human interactor
and an AI agent or agents that relies on equal contributions
to the developing story. This article focuses on one of the
major problems with co-creative systems and interactive
narratives in general: how the human user interacts with
the story world in an immersive fashion.
While intelligent agents in interactive narratives can
interact with humans through expressive embodied
representations, human interactors have no way to
reciprocate with embodied communication of their own.
Embodied agents can move, portray facial expressions, and
26
interactors and AI agents communicate with each other
through their performance rather than through behind-thescenes communication. Thus, the physical and diegetic
qualities of improv make it an ideal domain to study in
order to inform technological approaches to embodied cocreation.
This paper presents a system for combining improv
acting with full-body motions to support human-AI cocreation of interactive narratives. Our system translates a
human interactor’s movements into actions that an AI
improviser can respond to in the context of an emergent
narrative. We present a framework for human interaction
with an AI improviser for beginning an improvised
narrative with interaction mediated through an intelligent
user avatar. This approach uses human gestural input to
contribute part, but not all, of an intelligent avatar’s
behavior. We also present a case study of trained
improvisers interacting with the system, which informs
challenges and design goals for embodied co-creation.
While joint human-AI agents have been employed in
interactive narrative (Hayes-Roth, Brownston, and van
Gent 1995; Brisson, Dias, and Paiva 2007), this is the first
system to do so with full-body gestures.
space as the AI improviser. It gives the human interactor
an embodied presence in the scene and shows how the
Kinect senses their motion. This feedback can help the user
understand the avatar’s interpretations and adjust their
movements to accommodate the sensor’s limitations. The
avatar reasons about the human’s intentions for the scene
and creates its own mental model. Future implementations
of the system will follow the Moving Bodies interaction
metaphor with both the AI improviser and the intelligent
avatar producing dialogue based on their models of the
scene. The current implementation omits dialogue in favor
of studying gestural interaction with improvisational agents
alone.
The user stands before a large screen where the AI
improviser and the intelligent avatar are displayed on a
stage (Figure 1). The user faces the screen while standing
about four to ten feet away from a Microsoft Kinect below
the screen (i.e. within the Kinect’s sensor range). Both the
intelligent avatar and the AI improviser are shown as twodimensional animated characters. These simplified
representations map motion data from the Kinect directly
onto the characters’ animations. The system does not
currently support animated facial expressions, though such
animations may be supported in future iterations.
The user performs a motion to begin a scene, such as
putting one fist on top of the other and moving their hands
from side to side. The Kinect sends the motion data to the
intelligent avatar and the AI improviser, who each interpret
the motion as an action. The avatar displays the user’s
motion while it reasons about what character and joint
activity the user may be portraying. Here, the avatar may
identify the user’s motion as, for example, the action
sweeping and reason that they are portraying the character
shopkeeper. The AI improviser reasons about the user’s
Gestural Narrative Co-creation
Interactive narratives have used embodied gestural
interaction in stories with established characters (Cavazza
et al. 2004) or an established dramatic arc (Dow et al.
2007), though not for narrative co-creation. We enable
gestural interaction in our architecture, with co-creativity
as a goal, by combining elements from two real world
improv games, Three Line Scene and Moving Bodies
(Magerko, Dohogne, and DeLeon, 2011). In Three Line
Scene, actors establish the platform of a scene (i.e. the
scene’s characters, their relationship, their location, and the
joint activity they are participating in (Sawyer 2003)) in
three lines of dialogue. Improvisers establish the platform
at the beginning of a scene to create context before
exploring a narrative arc. We focus only on the character
and joint activity aspects of a platform as a simplification
of the knowledge space involved, since location and
relationship are often established implicitly or omitted
(Sawyer 2003). In Moving Bodies, improvisers provide
dialogue for a scene as usual, but let other people (typically
audience members) provide their movements. The mover
poses the improviser as if controlling a puppet while the
improviser creates dialogue based on their interpretations
of these poses.
In our system, the human interactor and an AI
improviser improvise a pantomimed three-line scene. An
intelligent avatar uses motion data from a Microsoft Kinect
sensor to represent the human’s motions in the same virtual
Figure 1. Human (upper left) controlling the intelligent avatar
(middle) while performing a three-line scene with an AI
improviser (right).
27
character and joint activity as well as its own character.
Here it may reason that the user is sweeping as part of the
joint activity cleaning. The AI improviser then chooses an
action to present (e.g. wiping tables as part of the joint
activity cleaning). It presents this motion to the user
through an animation. This animation would show the
improviser moving one hand at waist height repeatedly in a
circle. Then it is the user’s turn to present another motion
to the scene.
relating 59 actions, 11 characters, and 12 joint activities.
We chose the actions, characters, and joint activities based
on analysis of sixteen Western genre films and television
shows. Mechanical Turk allows us to survey a large
number of people and gather multiple data points for each
pair of associated elements. We can represent real-world
variability in human opinions by assigning the AI
improviser and the avatar different values for the same pair
of associated elements. Forcing the agent and avatar to
have slightly different background knowledge prevents
them from agreeing on aspects of the scene automatically,
making their interactions more like two improvisers
negotiating their mental models.
AI-Based Improviser and Avatar
The AI improviser and the intelligent avatar draw on the
same reasoning processes to understand how the human
interactor contributes to the scene. They utilize background
knowledge about a specific domain to make inferences
about the platform. They incorporate the human’s motions
into their reasoning based on joint position data from the
Microsoft Kinect sensor. Additionally, the AI improviser
reasons about how to contribute presentations to the scene
with motions. The intelligence given to the intelligent
avatar is intended to provide additional language
capabilities in future iterations of the architecture. If the
avatar can make reasonable interpretations of user gestural
input based on a) background knowledge and b) context
given from the scene, then it can add reasonable dialogue
utterances much like an improviser would that is
performing the Moving Bodies improv game.
Interpreting Motions as Actions
When the human interactor presents a motion to the scene,
both the intelligent avatar and the AI improviser need to
interpret the Kinect data for the motion as a semantic
action. Motions do not contain semantic information on
their own; the same motion presented in different contexts
can portray several actions (Kendon 2004). For example, if
an actor puts their fists together, brings them up to their
shoulders, and swings them in front of them, they could be
pantomiming swinging a baseball bat or chopping a tree
with an axe, depending on the scene.
The motion data from the Microsoft Kinect consists of
3D coordinates for joint and limb positions, which we
project to 2D and use to animate the on-screen characters.
The coordinates are evaluated as “signals,” which are
mathematical representations of specific joint angles,
relative joint positions, or changes over time. Signals can
be simple (joint angles and positions at one time) or
temporal (joint angles and positions varying across time).
For example, the arms crossed simple signal detects
whether each hand is positioned close to the other arm’s
elbow. The punching temporal signal, in contrast, detects
whether either arm rapidly extends and returns. Performing
this motion too slowly does not trigger the punching signal.
Temporal signals are evaluated over the span of a user’s
turn. A turn is considered complete when the human
interactor has either been out of a neutral stance for a set
period of time (currently two seconds) or returns to a
neutral stance after the system identifies at least one
candidate action. A neutral stance is defined as standing
still with feet together and hands at one’s sides.
Actions are defined as sets of positive and negative
signals. If the Kinect data satisfies all positive signals for
an action, it becomes a candidate. However, if the data
triggers any negative signals for that action, it is removed
as a candidate. When the user’s turn ends, the intelligent
agent and AI improviser select an action from the
candidate interpretations based on their mental model of
the scene (O’neill, et. al. 2011).
Background Knowledge
Both the AI improviser and the intelligent avatar reason
about how an improviser (human or AI) contributes to a
scene based on the fuzzy inference agent framework
described in (Magerko, Dohogne, DeLeon 2011). This
framework describes how agents can reason about
characters and joint activities in a scene based on motion
and dialogue inputs to infer fuzzy concepts and produce
perfomative outputs. While both the avatar and AI
improviser rely on procedural definitions of how to
negotiate a platform, they also need background
knowledge about a story world for those procedures to
operate on. The contents of this background knowledge
base can refer to any narrative domain; we have confined
our narrative domain, called TinyWest, to a set of actions,
characters, and joint activities associated with stories in the
Old West. The elements of the Old West genre have been
codified into generally recognizable stereotypes so that we
can assume that most people have a similar understanding
of these elements (e.g. most people have similar ideas
about what a cowboy is and does).
We collected crowdsourced data through Amazon’s
Mechanical Turk to find degree of association values (i.e.
bi-directional fuzzy relations) (O’Neill, et al. 2011)
28
Deciding on a Motion to Present
AI interpreted the improviser’s motion as. These two
conditions helped us study how displaying the AI
improviser’s interpretations of the user’s movements
affects the user’s understanding of and engagement with
the scene. The improvisers filled out a questionnaire about
their impressions after each scene. Part way through the
questionnaire, we revealed the AI improviser’s
interpretation of characters and joint activity in the scene
and asked the human to evaluate these interpretations.
After the second questionnaire, we conducted a brief,
structured interview with the participant.
Interpretation Feedback. The improvisers responded
to the on-screen interpretations with mixed feedback. Two
improvisers found that seeing the interpretation helped
them direct the scene. Improviser 3 considered changing
what he was doing to align with the AI’s interpretation.
Improviser 2 rated the interpretations as helpful (4 on a
Likert scale from 1 to 5, where 1 is “not at all helpful” and
5 is “extremely helpful”). This improviser thought that the
AI’s models of the scenes were at least moderately
reasonable in both scenes (at least 3 on a Likert scale from
1 to 5, where 1 is “completely unreasonable (seems
random)” and 5 is “completely reasonable (interpretations
make sense)”), the only improviser to do so.
In contrast, the displayed interpretations sometimes
confused the human improvisers, especially when the
interpretations diverged from their own. When Improviser
3 encountered a divergent interpretation, he said it “took
me out of the process of building a scene.” He described it
as “jarring that they’re putting a label, and that’s not at all
what I thought.” Improviser 1 reported feeling unsure how
to react upon seeing that the AI improviser’s interpretation
of her motion differed from her own. If she encountered
such a divergence in a scene with another human
improviser, she would have accepted their interpretation
and built off of that, a practice called “yes-and”-ing
(Johnstone 1999). Instead, she chose to follow her original
interpretation. This fits well with our understanding of
how ambiguity is used on stage (Magerko, Dohogne, and
DeLeon, 2011); improvisers rely on ambiguity in a
performance to allow for multiple possible interpretations
of the scene elements as they work to create a shared
mental model.
Improvisers rated the reasonableness of the AI
improviser’s interpretations higher when they received no
feedback about how it interpreted their motions. That is,
when the human improvisers judged the AI’s
interpretations of their motions only by observing its
motions, the humans considered the AI’s interpretations
more reasonable than when they saw explicit descriptions
of its interpretations. Three out of four human improvisers
rated the AI improviser’s mental model interpretations as
random (1 on a Likert scale from 1 to 5, where 1 is
After the AI improviser interprets the user’s motion and
reasons about how it contributes to the scene, the AI
improviser must select an action to present and a motion to
present it with. The AI improviser selects an action to
present in a similar way to how it reasons about characters
and joint activities (O’neill, et al., 2011). Once the AI
improviser selects a motion, its on-screen character plays
the corresponding animation. In our initial authoring, we
assigned one motion animation to each action. We
recorded motion data from the Kinect of one author
performing his interpretation of each motion. The author
posed in several successive keyframe poses for each
motion while the Kinect captured his body position at three
second intervals. Posing for keyframes rather than
recording continuous movements gave the author greater
control over which frames of motion the Kinect captured.
We took this approach to animation authoring for two
reasons: 1) by using an animation system of our own,
rather than an outside tool, we were able to test and iterate
on the signals and gesture recognition functionality
designed for the user character, and 2) this approach
simplified and accelerated animation work, which was
necessary since many animations were required and the
authors are not trained in character animation.
User Studies
Our system evaluation focused on case studies of how
expert improvisers interact with prototypes of our AI
improviser and intelligent avatar. We studied improvisers
from a local improv troupe as an initial case study of how
gesture can be used as a main interface paradigm for
interactive narrative.
We observed human improvisers playing our pantomime
version of Three-Line Scene, where the human and AI
improvisers attempt to establish the platform entirely
through pantomime. Our recruitment with a local theatre
yielded four improviser participants (one female, three
male) . Instead of ending the scene after three lines of
dialogue, the scenes ended after three motion exchanges
between the human and AI improvisers. This resulted in
six motions per scene, three from the human and three
from the AI improviser. The human contributed the first
motion to each scene and waited for the AI to respond. The
avatar still processed the human improviser’s motions and
reasoned about actions, characters, and joint activities.
After a practice scene to introduce them to the system,
each human improviser performed two scenes with the AI
improviser in the Tiny West domain. In the first scene, the
human received no feedback regarding the AI’s
interpretation of their motion. In the second, the system
displayed a one or two word description of the action the
29
“completely unreasonable (seems random)” and 5 is
“completely reasonable (interpretations make sense)”).
AI Improviser Movements. Three out of the four
improvisers
reported
difficulties
observing
and
understanding the AI improviser’s motions. Improviser 1
reported, “I don’t know what happened on their [the
agent’s] part,” claiming that the agent’s motions looked
like “just a bunch of movements.” Improviser 4 described
the agent’s movements as “random,” “undefined,” and
“difficult to interpret.” Improvisers 1 and 2 both described
the AI improviser’s motions as “fast.” They may have been
referring to the animations being short in duration or the AI
improviser presenting while they were still performing.
Improviser 2 said, “Normally, I’d let things build a little bit
to kind of feel out where I had a scene,” indicating that he
was still in the process of performing motions for his turn
when the agent presented. (Like the improvisers in the
previous study, these improvisers typically performed
multiple discrete motions in one turn.) Improviser 4
elaborated on the difficulties of turn-taking, saying, “We
either need to be able to engage simultaneously or engage
very, very distinctly. This felt like it was neither of those.”
Improviser 1 suggested that letting the AI start the scene
would ease the turn-taking at the beginning since the
human is better equipped to interpret motions and add to a
narrative than the artificial system. This result may point
to the need to work with improvisers in the future who are
trained in improvisational pantomime and the need to
jointly develop a gestural discourse vocabulary.
Differences from Improvising with Humans. The
improvisers noted a marked difference between interacting
with this system and their traditional improv experience.
Improviser 1, as noted above, seemed unsure whether to
interact with the AI as she would with a human improviser.
She was unsure how well the AI improviser would be able
to interpret her motions, saying, “I don’t know that it’s
intelligent enough to really know what I’m doing.”
Improviser 3 expressed a similar uncertainty and suggested
letting interactors “test ahead of time different gestures to
see what the system recognizes.” The human improvisers
could not treat the agent as an equal in their improvisations
because they could not assume a particular level of shared
knowledge. Additionally, the humans still felt like they
lacked non-verbal communication modalities even though
they could communicate with gestures. Improviser 4
specifically mentioned a lack of eye contact, which he
could use with another human improviser to establish that
they were both “on the same page.”
Preference for Face-to-Face Interaction. We initially
thought that displaying the improvisers’ actions on-screen
with the avatar would help them understand how the AI
improviser interpreted their movements. Three out of the
four improvisers said they would prefer a face-to-face
interaction with the agent rather than the “stage-view” that
showed both the AI improviser and the improviser’s
avatar. Improviser 1 missed presentations from the AI
improviser because she attended too much to the avatar. A
face-to-face display would immerse improvisers in the
scene more since the display would more closely mirror
their on-stage experience. Improviser 3, however, felt that
the stage-view “puts you in the scene.” He found that
seeing himself moving was helpful in a way that seeing
“just what your character sees” would not be.
Discussion
Our studies of expert improvisers interacting with
improvisational agents through full-body gestures indicate
several goals and challenges for future interaction designs,
encompassing both the AI improviser’s background
knowledge and the way that the scene is displayed to the
interactor. (The small sample size of our case study
prevents us from making broader generalizations.)
Balancing an improviser's traditional experience with the
need to present clear information about the AI's
contribution to the scene presents the main challenge of
designing embodied co-creative interaction with intelligent
agents for improvisers. While the visual display should
mimic traditional improv as much as possible to create an
immersive experience, other aspects of interaction need to
accommodate the AI improviser’s shortcomings to support
the human’s agency and clarity in the scene. Designing
experiences for people without improvisation training
should focus more on clearer presentations than
faithfulness to traditional improv. A visual display that
mirrors a traditional improvisation experience as much as
possible will create the best sense of immersion for a
human improviser. The screen should not show
interpretations of the improviser’s movements, since
seeing explicit, non-diegetic interpretations of divergences
reduced the human’s sense of immersion. Not having this
feedback will help human improvisers treat the AI
improviser more like a fellow human improviser. If the
human improviser is more comfortable with the AI
improviser, their interactions will feel more natural and
immersive. While we initially thought giving the user an
embodied presence in the same virtual space as the AI
improviser would make interaction more understandable,
this full stage view differs too much from the improviser’s
typical experience. Interaction with the AI improviser
should be presented as face-to-face, as improvisers would
interact this way on-stage. A transparent outline of the
human improviser’s motions in the foreground may be
necessary to represent the human in the virtual space in
future work when the intelligent avatar produces dialogue
on the human’s behalf as in Moving Bodies.
30
Hartholt, A., J. Gratch, and L. Weiss. 2009. At the virtual
frontier: Introducing Gunslinger, a multi-character, mixed-reality,
story-driven experience. In Intelligent Virtual Agents, 500–501.
Hayes-Roth, B., L. Brownston, and R. van Gent. 1995.
Multiagent collaboration in directed improvisation. In
Proceedings of the First International Conference on Multi-Agent
Systems (ICMAS-95), 148–154.
Johnstone, K. 1999. Impro for storytellers. Routledge/Theatre
Arts Books.
Laban, R., and L. Ullmann. 1971. “The mastery of movement.”
Magerko, B., and J. E. Laird. 2003. Building an interactive drama
architecture. In First International Conference on Technologies
for Interactive Digital Storytelling and Entertainment, 226–237.
Magerko, B., and M. Riedl. 2008. “What Happens Next?: Toward
an Empirical Investigation of Improvisational Theatre.” 5th
International Joint Workshop on Computational Creativity.
Magerko, B., W. Manzoul, M. Riedl, A. Baumer, D. Fuller, K.
Luther, and C. Pearce. 2009. An Empirical Study of Cognition
and Theatrical Improvisation. In Proceeding of the Seventh ACM
Conference on Creativity and Cognition, 117–126.
Magerko, B., Dohogne, P. and DeLeon, C. 2011. Employing
Fuzzy Concepts for Digital Improvisational Theatre. In the
Proceedings of the Seventh Annual AI and Digital Interactive
Entertainment Conference, Palo Alto, CA.
Mateas, M., and A. Stern. 2003. Integrating plot, character and
natural language processing in the interactive drama Façade. In
Proceedings of the 1st International Conference on Technologies
for Interactive Digital Storytelling and Entertainment (TIDSE03).
McCoy, J., M. Treanor, B. Samuel, N. Wardrip, and M. Mateas.
2011. Comme il faut: A system for authoring playable social
models. In Seventh Artificial Intelligence and Interactive Digital
Entertainment Conference.
O’Neill, B., Piplica, A., Fuller, D., and Magerko, B. 2011. A
Knowledge-Based
Framework
for
the
Collaborative
Improvisation of Scene Introductions. In the Proceedings of the
4th International Conference on Interactive Digital Storytelling,
Vancouver, Canada.
Riedl, M., and A. Stern. 2006. “Believable agents and intelligent
story adaptation for interactive storytelling.” Technologies for
Interactive Digital Storytelling and Entertainment: 1–12.
Sawyer, R. K. 2003. Improvised dialogues: Emergence and
creativity in conversation. Ablex Publishing Corporation.
Tanenbaum, J., and K. Tanenbaum. 2008. “Improvisation and
performance as models for interacting with stories.” Interactive
Storytelling: 250–263.
Young, R.M. 2001. An overview of the mimesis architecture:
Integrating intelligent narrative control into an existing gaming
environment. In The Working Notes of the AAAI Spring
Symposium on Artificial Intelligence and Interactive
Entertainment, 78–81.
Zhang, L., M. Gillies, J. Barnden, R. Hendley, M. Lee, and A.
Wallington. 2007. “Affect detection and an automated
improvisational ai actor in e-drama.” Artifical Intelligence for
Human Computing: 339–358.
The human improvisers’ perception that their
interactions with the AI improviser were random arose
from unclear presentations from the AI. This perceived
randomness in turn inhibited the human improvisers’ sense
of agency because they felt that their presentations did not
cause a sensible response from the AI improviser. While
ambiguity and unexpected presentations are a natural part
of improv, presentations should still make sense within the
context of a scene. One factor contributing to this may
have been the lack of more subtle non-verbal
communication from the AI, for example in the form of
facial expressions, eye movements, or small changes in
posture, that could provide clues to both the AI’s
interpretation of the improviser’s action and the intention
in responding. Improving the quality of the AI
improviser’s motion animations so that they are as life-like
as possible will alleviate the perception of randomness and
improve both clarity and agency. Chained and repeated
motions in a human improviser’s presentations made it
harder for the AI improviser to moderate turn-taking. To
ease this difficulty and to increase the human improviser’s
sense of agency, the user should explicitly indicate when
their turn ends. The human improviser will clearly know
when to attend to the AI improviser. An invented hand
signal (which the improviser will not use in their natural
presentations) can communicate the end of a turn.
Although human-moderated turn-taking departs from the
continuous flow of presentations in traditional improv, this
sacrifice enhances the clarity of presentations. These
results indicate the need for studies in the future that
remove the AI from the system and rely on a Wizard of Oz
technique to clearly focus on the gestural interactions
without distraction from how intelligent the AI may or may
not be.
References
Aylett, R. S., S. Louchart, J. Dias, A. Paiva, and M. Vala. 2005.
FearNot!-an experiment in emergent narrative. In Intelligent
Virtual Agents, 305–316.
Brisson, A., J. Dias, and A. Paiva. 2007. From Chinese Shadows
to Interactive Shadows: Building a storytelling application with
autonomous shadows. In Proceedings of the Agent-based Systems
for Human Learning Workshop (ABSHL) at Autonomous Agents
and Multiagent Systems (AAMAS), 15–21.
Cavazza, M., F. Charles, S.J. Mead, O. Martin, X. Marichal, and
A. Nandi. 2004. “Multimodal acting in mixed reality interactive
storytelling.” Multimedia, IEEE 11 (3): 30–39.
Dow, S., M. Mehta, E. Harmon, B. MacIntyre, and M. Mateas.
2007. Presence and engagement in an interactive drama. In CHICONFERENCE-, 2:1475.
El-Nasr, M., K. Isbister, J. Ventrella, B. Aghabeigi, C. Hash, M.
Erfani, J. Morie, and L. Bishko. 2011. “Body buddies: social
signaling through puppeteering.” Virtual and Mixed RealitySystems and Applications: 279–288.
31
32